Face Recognition Based on Point Cloud Data Captured by Low-cost mmWave Radar Sensors

The image based face recognition is a well-established technology nowadays. However, commercially available cameras generally perform poorly in environments with low visibility such as in dark night or foggy weather. In addition, reliable privacy protection to avoid potential data leakage poses quite a challenge for resource-constrained Internet of Things (IoT) devices at edge, particularly for cameras producing visual data. As a promising alternative, we propose to use commercial low-cost mmWave sensors to assist or perform completely the face recognition task. In our approach, we directly leverage the point cloud data captured by an off-shelf mmWave radar sensor to train a neural network modified based on PointNet. Based on our modified PointNet architecture, we are able to achieve an overall accuracy of 98.69%, with much lower computation and data bandwidth requirement as compared with the image based approach. We believe our work serves as a great example of the potential of using mmWave sensors for a richer understanding of the sensing environments.